🤖 AI Summary
This work addresses the lack of controllable refusal capability in large language models (LLMs). We propose the first unified framework for uncertainty calibration and risk-controlled rejection tailored to API-based black-box LLMs—requiring no fine-tuning and offering distribution-free theoretical guarantees. Methodologically, it integrates heterogeneous uncertainty signals—including sequence likelihood, self-consistency dispersion, retrieval compatibility, and tool feedback—into a lightweight calibration via temperature scaling and adaptive scoring, then enforces principled rejection using conformal risk control under user-specified error budgets. Key innovations include fine-grained factual alignment and interpretable refusal. Experiments across short-form QA, code generation, and retrieval-augmented long-text generation demonstrate substantial improvements over entropy- and logit-threshold baselines: lower calibration error, superior area under the risk–coverage curve, and higher coverage at fixed risk levels.
📝 Abstract
Deployed language models must decide not only what to answer but also when not to answer. We present UniCR, a unified framework that turns heterogeneous uncertainty evidence including sequence likelihoods, self-consistency dispersion, retrieval compatibility, and tool or verifier feedback into a calibrated probability of correctness and then enforces a user-specified error budget via principled refusal. UniCR learns a lightweight calibration head with temperature scaling and proper scoring, supports API-only models through black-box features, and offers distribution-free guarantees using conformal risk control. For long-form generation, we align confidence with semantic fidelity by supervising on atomic factuality scores derived from retrieved evidence, reducing confident hallucinations while preserving coverage. Experiments on short-form QA, code generation with execution tests, and retrieval-augmented long-form QA show consistent improvements in calibration metrics, lower area under the risk-coverage curve, and higher coverage at fixed risk compared to entropy or logit thresholds, post-hoc calibrators, and end-to-end selective baselines. Analyses reveal that evidence contradiction, semantic dispersion, and tool inconsistency are the dominant drivers of abstention, yielding informative user-facing refusal messages. The result is a portable recipe of evidence fusion to calibrated probability to risk-controlled decision that improves trustworthiness without fine-tuning the base model and remains valid under distribution shift.